Can AI Learn Physics by Watching Videos? The Answer Is More Complicated Than It Sounds
Today’s AI stories share a quiet theme: the gap between what these systems appear to do and what they actually understand. From robots trying to grasp the physical world, to questions about whether an AI might feel pain, to a state government hitting the brakes on the infrastructure powering it all — the limits of AI are in focus right now.
AI Systems That “Watch” Videos to Learn How the World Works
World models — AI systems trained to predict what happens next in a video — are having a moment in research labs, and the stakes are higher than they might seem. Rather than learning from written instructions or labeled data, these systems watch thousands of video clips and try to anticipate the next frame. The idea is that if an AI can accurately predict what a falling glass will do, or how a robot arm should move to pick something up, it might develop a working understanding of physics itself.
Think of it like learning to drive by watching dashcam footage instead of sitting in a car. You might get good at predicting what other drivers will do, but whether you truly understand momentum and friction is another question entirely.
That gap is exactly what experts are debating, according to this detailed breakdown from Ars Technica. Some researchers believe world models could eventually let robots train in realistic simulations before ever touching real equipment. Others argue the systems are pattern-matching on visual information without grasping the underlying rules.
For regular people, the practical question is whether the next generation of robots — in warehouses, hospitals, or homes — will be genuinely reliable. A robot that learned physics by watching videos might handle familiar situations well and fail badly in ones it hasn’t seen.
Why this matters: Better world models could make robots far more capable. But if those models have hidden gaps in their understanding, the failures could be hard to predict and harder to catch.
“World models learn to predict what happens next in videos by watching examples.”
New York Just Put a Pause on New Data Centers. Here’s What That Signals.
New York has reportedly become the first U.S. state to implement a temporary moratorium — a formal pause — on permits for new large-scale data centers, according to The Verge. The freeze lasts up to one year. The governor is also reportedly weighing legislation that would place stricter limits on where future data centers can be built and how they operate.
Data centers are the physical buildings that house the servers running AI tools, cloud services, and streaming platforms. Large ones consume enormous amounts of electricity and water for cooling. New York’s move reflects growing tension between the tech industry’s appetite for computing power and local concerns about energy grids and environmental impact.
For people living near proposed data center sites, this matters immediately. Several communities across the country have raised concerns about noise, water use, and strain on local power supplies. New York’s pause gives policymakers time to write clearer rules before more projects break ground.
Why this matters: One state moving first often sets a template others follow. If New York’s regulations prove workable, expect other states to use them as a starting point.
“First state to enact a data center moratorium, up to one year.”
Anthropic Asked Whether Its AI Might Be Suffering. The Research Is Unsettling.
Anthropic, the company behind the Claude AI assistant, has published research exploring the inner workings of its models — including the unsettling question of whether those models might experience something like pain or suffering. Interpretability research, which is the field of trying to understand what is actually happening inside an AI system rather than just observing its outputs, hasn’t tackled questions this philosophical before. MIT Technology Review has a close look at what the work does and doesn’t show.
Anthropic’s researchers aren’t claiming their models are conscious. The work reportedly tries to identify internal patterns that might correlate with concepts like discomfort, and asks whether those patterns influence the model’s behavior. It’s the difference between a thermostat registering heat and a person feeling hot.
For everyday users, this research matters in a different way than the headline suggests. It’s less about robot rights and more about trust: if AI companies can see inside their models more clearly, they can build systems that behave more consistently and catch problems before they reach users.
Why this matters: The fact that a major AI lab is funding this kind of research signals that questions about AI behavior and internal states are moving from philosophy classes into engineering teams.
“Research exploring whether AI models might experience pain or suffering.”
Also Happening in AI
Ollama released version 0.32.0 of its popular tool for running AI models locally on your own computer, adding an interactive agent feature for coding and conversation. Meanwhile, Google and India’s AIM program launched ATL Saathi, a Gemini-powered tool designed to help teachers support student innovators across the country. On the funding side, AI video generation startup PixVerse raised $439 million at a valuation above $2 billion, per TechCrunch — a sign that investor appetite for video AI remains strong. Nous Research, which builds AI agents for automating tasks, is reportedly in talks for new funding at a $1.5 billion valuation. Uber’s product chief also weighed in on the company’s push into hotels and robotaxis, while being careful to say Uber doesn’t want to expand into every corner of daily life.
What to Watch
The data center moratorium in New York and the Anthropic interpretability research point to the same underlying tension: AI is scaling fast, and governance — both technical and political — is struggling to keep pace. Watch whether other states follow New York’s lead before the one-year pause expires, and keep an eye on whether Anthropic’s interpretability findings prompt other major labs to publish similar internal research. If they do, 2026 may be the year the industry stopped treating AI’s inner workings as a black box.